Azure ML Docker Examples

Some examples of working with docker in Azure (ML).

Example: Pull an image from MCR and run locally

To pull this image

mcr.microsoft.com/azureml/bert:pretrain-openmpi3.1.2-cuda10.0-cudnn7-ubuntu16.04

and run it locally, do:

$ docker pull mcr.microsoft.com/azureml/bert:pretrain-openmpi3.1.2-cuda10.0-cudnn7-ubuntu16.04
$ docker run -it mcr.microsoft.com/azureml/bert:pretrain-openmpi3.1.2-cuda10.0-cudnn7-ubuntu16.04
root@99ab01080c61:/#

Example: Dockerfiles & Azure ML

Step 1 - Write dockerfile

Start with a dockerfile

# ./Dockerfile
FROM mcr.microsoft.com/azureml/base

RUN pip install --upgrade pip
RUN pip install azureml-core==1.7.0 azureml-sdk==1.7.0 scikit-learn==0.23.1

Step 2 - Azure Container Registry

Push this image to the Azure Container Registry associated to your Azure ML Workspace:

1. Find the name of your existing Azure Container Registries (there will be one created when you spin up an Azure ML workspace)

   az ml workspace show -w <myworkspace> -g <resourcegroup> --query containerRegistry

   You'll get something like this:

   /subscriptions/<subscription_id>/resourceGroups/<resource_group>/providers/
       Microsoft.ContainerRegistry/registries/<registry_name>

   Make a note of the <registry_name>.

   ??? failure "Error: Workspace not found"

   If you get an error that the workspace cannot be found you may need to set your subscription:
   
   ```bash
   # by name
   az account set --subscription "My Demos"
   
   # or by id
   az account set --subscription 42ae47bd-b19b-42c1-b0b9-19fd5be9d51b
   ```

• Login:

  az acr login --name <registry_name>

• Build your image in the cloud. Navigate to your dockerfile and run this to create and push the image into your Azure Container Registry:

  az acr build --image myimage:v1 --registry <registry_name> --file Dockerfile .

Step 3 - Reference the container in your Azure ML launch script

In this example we use the Azure ML python sdk to kick off the job. We need to reference our image here.

# get workspace
ws = Workspace.get(...)

# Add training script to run config
src = ScriptRunConfig(
    source_directory=".",
    script="script.py",
)

# to run on local compute
src.run_config.environment.python.user_managed_dependencies = True
src.run_config.environment.docker.base_image = <azure-container-registry-image-name>

# Create experiment
exp = Experiment(...)

# Submit run
run = exp.submit(src)

To be explicit, when I ran the command

az acr build --image amsaied/async:v1 --registry amsaiedamlws8d0e0159 --file dockerfile .

Within the output you'll see this:

image:
    registry: amsaiedamlws8d0e0159.azurecr.io
    repository: amsaied/async
    tag: v1

which I can reference like this

amsaiedamlws8d0e0159.azurecr.io/amsaied/async:v1

for example

src.run_config.environment.python.user_managed_dependencies = True
src.run_config.environment.docker.base_image = 'amsaiedamlws8d0e0159.azurecr.io/amsaied/async:v1'

"Warning: AML + WSL != Docker"

As of 2020/06/12, Azure ML does not support running docker locally from WSL.